#24 · ML / AI methods

Statistical methods for microbiome inference under compositionality and confounding

The dominant thread is statistical methodology for microbiome data, where compositional structure and confounding complicate inference. A phylogeny-guided Bayesian shrinkage method (PGL) addresses ASV-level feature selection via phylogenetic borrowing, and a benchmark of 71 single-cohort and 63 multi-cohort confounder-aware differential abundance strategies assesses methods under simulation and real data. A high-dimensional structural equation framework links large microbiome cohorts with smaller paired microbiome-metabolome data to build disease pathways. Benchmarking of 18 co-occurrence network methods on synthetic 16S count data with known interactions finds validity limited, and graph neural networks on individual association networks identify a co-abundant Bacteroidales module driving IBD prediction. The shared element is explicit handling of uncertainty and ground-truth validation, though the items differ in analytical aim.